Discussion in the previous post’s comments has come around to what we think we know and what we might be able to infer about the genetics of Uralic speakers. I have no direct data on this at my disposal (nor would I know off the cuff what to do with it if I did), but once again I find that already staring at published results seems to show a lot of details that would allow also closer discussion.
The most popular presentation of results on Uralic overall genetics might be still the 2018 paper “Genes reveal traces of common recent demographic history (…)” by the Tambets et al. group (surely it is by now the most influential, at least). This demonstrated, firstly, a broad west–east cline structure with offshoots to the south (Hungarian) and north (Samic), which may have been about as expected. But more interesting may have been their ADMIXTURE charts in Fig. 3A, that revealed at 9-component resolution a component (in magenta in their figure) that seems to correlate surprizingly well with Uralic-speaking peoples.
The match is still far from perfect. This component seems to be basically absent in Hungarians, while being well present in Volga and South Siberian Turkic peoples, outright a majority in Kets, and trickling marginally also into e.g. Russians, Swedes and Poles in the west (much less so in Latvians), Mongolians in the east. Its peak representation within the study is at >90% in Khanty & in a distinct subgroup of Mansi. Some Selkups also show a high proportion. Elsewhere it is mostly 30–40% though, usually in primary competition with another ancestry component (in navy) that is well spread thru Europe and reaches >90% in Latvians, almost as much in Lithuanians and Estonians. The lowest proportion of magenta in Uralic speakers occurs in Nganasans, whose main ancestry component at this resolution (in orange) is in common with all Altaic-speaking populations (> 80% in Yakuts, Evenkis, Evens), well present also in Yukaghirs and Kets and with a clear trace in Japanese and a subgroup of Han. All these complications have left me with the feeling that trying to ascribe this ancestry as being specifically spread by Proto-Uralic speakers may be too optimistic. Most pessimistically, it could represent instead an already pre-Uralic genetic cline, or assimilation to one — especially under current theories where Uralic spread (at least) to the west essentially within a metal trade network, thus perhaps as a trade language, at which point in time there may then not even have been such a thing as a single “Proto-Uralic-speaking people” anymore. Probably further back in time we could find something fitting this definition (and contrary to my last post, still most likely as the actual majority ancestors of at least one actual Uralic-speaking group), but if people A teach their language to people B who teach it to people C and D, who teach it to…, we might quickly end up chasing false leads about genetic markers.
Tambets et al. highlight also the 11-component resolution, where a new component appears (in light blue) that captures both the previously magenta + navy ancestry of the Swedish Sami, ending up at 98% among them!, overtakes magenta also among Kola Sami and Finnic speakers (but interestingly not North Russians) and appears in some 5%-ish proportions eastwards until the Mansi (and some 1%-ish trace presence all across Siberia). This remains without specific discussion in the paper, but has by now often come up in discussions with colleagues. My current understanding is that the ADMIXTURE algorithm in fact prefers to fit components to specific populations or individuals actually present in its data. And when a population has a distinct ancestry component not found in pure form (in this case we are likely dealing with the Paleo-Laplandic and/or Paleo-Lakelandic substrates in the Sami), it cannot reconstruct it. Instead a new component ends up representing this actual ancestry + a proportion of a reference group’s other compareable ancestry components, which then means even a small fragment of this “underlying” ancestry, together with other shared components, will now appear in larger size also elsewhere. So since the Sami still have some general-European ancestry, perhaps with drift or selection having emphasized some specific part of it, we find this reflected in a trickle of light blue also in e.g. Bulgarians or Hungarians. Since they also still have some eastern / Siberian ancestry, we find a similar trickle also in e.g. Udmurts (who have a good amount of general-European ancestry too) or Khanty. But then if this component is not directly representative of some old Proto-Sami or pre-Proto-Sami ancestry, and instead groups in also parts of other ancestry components that tend to occur together with one, why not also the same in other components? We also see that the equation 50% magenta + 50% navy = light blue, roughly valid for the Sami, does not hold in other populations. “Finnish light blue” is closer to one part magenta + two parts navy, and also leaves a little bit of magenta behind. “Veps light blue” looks more like the inverse: 60% magenta + 40% navy. “Komi light blue” and “Udmurt light blue” look to me to be taking in a dash of green, too. “Mari light blue”, looking closely, could involve a tiny reduction also in their yellow-coded ancestry (but also alternately, from the appearence of the light brown ancestry component in the chart).
Moreover I have not seen or heard it noted so far at all that their fuller data, provided as an appendix, includes also further turns of this sort. A very similar update seems to happen at the 14-component resolution: here a component in dark orange appears that captures the majority of the previously magenta and orange ancestry of the Nganasans. Other peoples who gain a large chunk of this end up being the Nenets and Selkups to the west, Evenkis, Evens and Oroqens to the east, and one specific Yakut individual and Koryak individual. Smaller amounts appear in e.g. Khanty, Tuvinians and in a sub-population of Dolgans, and a small sliver in all Japanese. Interestingly this component also, even at a low proportion, seems to completely replace light orange ancestry in several Uralic-speaking groups: besides the Nenets, who now end up as ca. 60% magenta, 40% dark orange, this also holds for the Khanty, Mansi and Komis. And also, not for Maris. Surely a sign that their orange ancestry is specifically recent from Turkic speakers, not ancient Siberian independently of general Uralic origin.
Should we assume this is, as I speculated might be the case for the Sami-affiliated component 11, in reflection of some lost substrate population that was closest of all to Nganasans? This I am much less sure of. The order in which components appear in ADMIXTURE results is clearly largest-first. Component 14 is found in only few peoples and is thus forced to be distinguished late from “general Altaic” ancestry. I must wonder is this might change if we pared the dataset down, making e.g. Nganasans a larger proportion of the data; and perhaps in a different order we’d find also find different split topology, such as 14 initially grouping with 9?
A second update to the component analysis of several Uralic-speaking groups is no. 16 in pink, likely representing no ancient ancestry at all but the formation by convergence of a specific secondary mixture: a “Volga region” component, appearing at 90% in Maris, very high also in Udmurts and Chuvashes. There is medium presence in Bashkirs, Tatars and Komi, smaller amounts appear in Russians, Mordvins and Mansi, and clear traces in all Turkic-speaking people and even in Poles. Interestingly, perhaps none in the Finnic peoples (but the appendix image is not high resolution enough for me to be sure of this).
Here we might move in the blog post from “reading” into “overreading”. Can we perhaps on closer looking work out what does the order and exact form of the splitting of ADMIXTURE components signify more specifically?
K=3. At this resolution we have three very broad components: a dark brown component that we might call sub-Saharan African or West African, at 100% in Mandenkas and Yorubas; a blue component at 100% in e.g. Basques and widely spread in Eurasia; and a yellow component at 100% in e.g. the Han and the Japanese. [1] Not highly useful for understanding anything about Uralic peoples: everyone on European side mostly blue with a small dash of yellow, Khanty and Samoyedics majority yellow with a clear piece of blue; though there are signs of other interesting patterns elsewhere, such as about 40% blue in Ethiopians, 5% brown in Georgians and Armenians.
K=4. This splits the blue component in two, into a broadly northern one in navy and a southern one keeping the medium-blue. No one ends up as purely either. Blue without navy appears, it seems to me, only in Ethiopians and Indians! The highest absolute amount of navy is in Estonians, Latvians, Russians at around 80%. Armenians and Georgians come out at the other end with 90% blue, 10% navy. The Swedish Sami, Nenets and Khanty (but no one further east) seem to end up with complete blue → navy conversion. So might e.g. Dolgans, Evenkis or Kets, but they have so little blue that I cannot be sure. Yakuts keep on a sliver of medium blue, Volga region Turkics follow median Uralic speakers in becoming majority-navy, other Turkic speakers maintain medium blue as the majority.
K=5. This splits off from yellow an orange component, that appears broadly across Siberia and/or Uralic speakers and in traces across Europe. The westernmost people who retain any consistent yellow are Bashkirs. Most Common Turkic speakers have a nearly even split. Chuvashes by contrast convert nearly all their yellow into orange, patterning here as western. Yellow remains in Indian control groups like Kannadigas. Han gain some 2% orange but remain the “type specimen” of yellow. The Japanese and a Han subgroup that also shows a small dash of blue and, later on, green (perhaps indicative of something like Manchu ancestry?) gain maybe 3–4%.
K=6. A light green component appears that is at its maximum in India, replacing both blue and orange here. Most Iranic peoples and neighbors (the Burusho and Brahui) go from about 90% blue / 10% orange (and a dash of other colors) to about 65% green / 35% blue, and seem to add back a trace amount of yellow. Green appears also in people from the Caucasus (at the expense of all of orange, blue and navy) and, less strongly, in Turkic and Uralic speakers. Further west, we see in Saudis, Yemenites and Palestinians a seemingly total conversion of both orange and navy, some decrease in blue too, for 15% green. Turks, Azeris, Kurds, “Iranians” (= Persians proper?) add an even more noticable 20% but still keep maybe 2-5% navy and orange. Orange traces in Europe mostly turn green, with the exception of Russians and Germans. In Moroccans and in Mzāb (a Berber group) navy now disappears, Sardinians go from 50% blue / 50% navy to 75% / 25%, and Basques from 33% / 66% to 50% / 50% — without any green appearing in any of these four. In Ethiopians a tiny sliver of green turns up. Altogether a confusing turn of components that seems to be mainly really about rebalancing something about the very vague blue / navy split.
By this point it seems navy is no more generally northern but rather northwestern in its distribution, still at its strongest in Balto-Slavic, Finnic and Samic speakers, with Germanics and Volga Region Uralics not far behind. I do not know what to make of its medium presence in Siberia and therearound, with fairly high navy : green rations in e.g. Kets, Selkups and Mongolis. Some individuals are probably though simply mixed half-Russian: there looks to be e.g. one Yukaghir, two Koryaks, at least three Chukchis who remain about 50% navy + medium-blue to the end of the analysis even after other components yet have been split off from these. One other “Chukchi” looks entirely Slavic; or possibly e.g. part Mordvin or German, but anyway not at all Siberian.
K=7. A tan “Beringian” component splits off from the general Siberian orange, at almost 100% in most Koryaks and Chukchi, a neat 50% also in the individuals with large blue and navy proportions, and remains definitive of these people until the end of the analysis. Only the Evens gain a noteworthy minority proportion, but the better-resolution data in the main paper shows that a small 1–2% sliver of this component still extends thruout Asia and into eastern and central Europe in e.g. Bulgarians or Poles or Finns. Not, however, into the Sami or Swedes.
K=8. The confusing green component splits further. The Indian half of it turns dark olive, at its highest proportion in “Malayans” (does this mean Malayalams? I would not guess Malays). A small portion of this appears in Iranians, Azeris, Turkmens and an even smaller trace in Syrians, Yemenese, Lezgins, but Near Easterners mostly keep light green. Pathans, Balochis etc. end up evently split or slightly more green than olive. In the Caucasus light green now becomes the majority over blue. Europe thru Siberia show only light green; which also seems to marginally increase at the expense of blue in e.g. French, Italians, Maris and also in Ethiopians. No effects on proportion of navy. An odd side effect might be also turning the sliver of eastern ancestry in Turks and Azeris now from orange to yellow, although the proportions of these across Siberia and Central Asia remain stable.
K=9. Enter the “Uralic” component, distributed as described above. This turns out to be mainly at the expense of orange and navy. Most Khantys remain 98% magenta from here to the end of the analysis. In peoples who were not included in the figure in the main paper, a small dash appears also in e.g. Iranians, Germans, Caucasians (Chechens and Lezgins in particular), Pathans, Burushos, and even Mzābs. In a few of these cases, detectable amounts of Proto-Uralic or even para-Uralic ancestry would be quite surprizing, I think.
K=10. A light brown component appears at the expense of primarily blue, secondarily dark brown, up to 80%+ in Mzābs. Replaces blue completely in Ethiopians and also replaces small traces of blue that up to now have been present in the third (and vague) deep African control group of “Bantus”. In Saudis this replaces dark brown completely but leaves a tiny sliver of blue. Outside the Near East this appears e.g. at 20% in Cypriots, 10% in Portuguese, Sardinians, Armenians (only a trace in Georgians), maybe <5% in Spaniards and across the Balkans. A few localized bits of dark brown in e.g. Sindhi or Balochi seem to remain. The new version of medium blue concurrently becomes more prevalent in West Europe at the expense of navy, rising e.g. from 25% to 50% in French, from 50% to 80% in Basque. Nothing changes in Siberia.
K=11. The light blue Swedish Sami-maximized component appears and remains at >98% in them until the end of the study. No apparent side effects beyond the Uralic/Nordic zone.
K=12. Light brown splits in two. The component that keeps the color grows to 100% in some Mzābs and remains there until the end of the study, is well-present also in Moroccans, and appears also in Portuguese and Spaniards (thus, at the western end of its range). Probably everywhere else previous light brown turns a pinkish brown. Balochi and Brahui seem to develop a larger proportion of brown, maybe at the expense of their previous blueish (either medium or navy) ancestry.
K=13. Olive splits in two and claims also some more of light green, whereever it was present. The component that keeps the olive color is now at 80%+ in Brahui (previously 60% green 40% olive) and becomes a more notable minority in e.g. Iranians, Tajiks and Lezgins. A light dirty green becomes the new South Indian majority and also ousts remaining yellow from here. The only clear inroads this makes beyond South Asia are into Pathans and Burusho, maybe a sliver in Tajiks and Uzbeks. Light green claims also yet more of blue from Georgians and Abkhazes, rising in them now to 90%.
K=14. The Nganasan-maximized component in dark orange, as noted above. As one side effect, the proportion of magenta in Kalmyks rises from marginal to maybe 10%, at the expense of light orange.
K=15. Another split from the generic-Siberian orange: a grey component that claims about 90% ancestry in Yakuts (aside from the one Nganasan-like, or even more closely, Even-like individual), 75% in Evenkis, 60% in Dolgans (who all now have about no orange); 50% in Yukaghirs and Evens (who retain 10% and 15% orange respectively); 5% in Buryats and Mongolians (who remain majority orange), traces in Turkic speakers further to the west. These colors, too, thus remain largely orthogonal to the linguistic division of Siberia. Light orange proportion in Han seems to rise slightly.
K=16. The Volga component in pink, as noted above, and yet another one orthogonal to linguistic classification. No side effects elsewhere that I can see on the proportions of magenta / green / navy that it replaces.
K=17. Some sort of a Levant component in dark grey enters, at 100% in some Palestinians but petering out to <5% in others; mostly thus replacing pink-brown and green. 5% also in Cypriots, only trace amounts in Jordanians and, oddly enough, Basques, Romanians and Gagauzes. I’m not sure if this has any presence anywhere else.
K=18. Here we gain a Sardinian component in purple, replacing almost all of their medium blue and also most of their pink-brown. Concurrently, Basque now becomes the unquestioned type specimen of medium-blue at maybe 98%, losing even its last traces of navy; its dark gray from the previous step seems to turn into purple, though! In the Middle East and southern Caucasus (Armenians, Georgians, Abkhazians) blue also mainly turns into purple. Italians, Romanians and Slavic speakers gain more purple than blue; more blue than purple at least in western Romance, Germanic, Iranic, Finnic, Mordvinic speakers and everyone in the North Caucasus, including the Turkophones (Balkars, Kumyks, Nogais).
K=19. We have one last change affecting Uralic speakers: a sky blue component appears at 80%+ in Ket and very variable among Selkups, anywhere from 90%+ to 10%. Smaller proportions appear mainly in South Siberian Turkic: Tuvans, Shors, Altais, Khakassians; as well as in Nenetses and in the non-Khanty-like part of Mansis (the Khantys still remain solid magenta). Possibly a dash in Uzbeks and Tajiks, though these appear by now as highly mixed populations with a little of almost every component (the former plurality orange, the latter plurality green + almost as much olive). Maybe same in Kyrgyzes and Kazakhs, who have a stronger orange plurality but still likewise a very diverse collection of components.
K=20. The last component of the study rearranges again the greenish colors: a new bright green hits a 90% among some of the Burushos, turns up in some proportion in most Iranics / Central Asians, has again a spot appearence among Lezgins and Chechens, and I think even at least one Syrian. Really nothing of interest for Uralic though.
My initial impression from this deep dive is not so much that I’ve definitely learned new things (though interestingly many known results of population genetics do appear here in a nutshell, such as Near Eastern connections of the Sardinians or the Ethiopians) as much as, first of all, that there would be likely benefits to re-doing this more detailed analysis with fewer “distractions”. Suppose we discard those peoples that end up generating their “own” components without much relevance for distinguishing Uralic speakers or their neighbors: that is the Mzābs and Moroccans (#12), South Indians (#8 and #13), Palestinians and Jordanians (#17), Burushos (#20), maybe one (but maybe not both) of the Sardinians & the Basques (#18). Do we still find the other components appearing in the same shape and same order? What six other components do we obtain instead if running the analysis to K=20?
We might also leave aside groups like the Japanese, Orcadians, Balkars, Abkhazians, Ossetians who don’t seem to add much of their own, to have Uralics and neighbors be a larger proportional part of the data (yes, Northwest Caucasians and Ossetians may have had contacts with migration-era Hungarians, but we can by now take it as known that modern Hungarians look genetically generically Slavic ~ Romanian). And since the algorithm seems to also look at the pre-given ethnicity bins, what would happen if we e.g. split in advance the Mansi or Kola Sami samples in their apparent subgroups? Or removed the apparent half-Russians from amongst many peoples of Siberia?
My second observation is that there’s a suspicious lack of any component that we could think of as even vaguely tracking most other larger language families: Indo-European (or its subgroups) or Turkic or Tungusic, without spilling in large percentage into their neighboring language families. Orange after the Chukotkan split in K=7 does maybe a tolerable job at tracking classical Ural–Altaic, but still keeps on also Yukaghir and Ket. And there is no hope at any of the resolutions here for telling apart e.g. an Abkhaz from a Balkar or a Persian from an Azeri. Perhaps the ability to still distinguish Finns and Swedes, or Khantys from everyone, is somewhat a happy accident really. Everywhere else Uralic-speaking people continue to look more similar to their neighbors than to their distant linguistic relatives.
Thirdly, towards the end of study we have also ended up with a fairly clear split of components into two sorts: those that have their “type specimens” and those that do not. Most of the former seem to be also very localized, though a few are not.
- Typified, widespread: medium blue (Basques), purple (Sardinians), yellow (Han).
- Typified, local: dark brown (sub-Saharan African), medium brown (Mzābs), dark grey (some Palestinians), light blue (Swedish Sami), magenta (Khantys), dirty green (some South Indians), sky blue (Kets + some Selkups), dark orange (Nganasans), tan (Chukotkans). The most borderline cases are pink (Maris; still with small bits of navy / green / magenta), grey (Yakuts; still with bits of, at least, dark orange) and medium green (Burushos; still with other green hues).
- No type specimen, widespread: pink-brown, navy, light green, dark green, orange.
The impression I get from the big picture is that these “components” are thus unlikely to represent any sort of recent unique markers. They seem to come mainly from already Paleolithic genetic variation, sometimes probably eventually approximating unique “poles” (yellow, medium-blue, maybe tan) but more often probably representing specific proportions of mixture, especially if that has been later on well preserved in some peoples. So in particular, even if component 9 had something to do with the spread of Uralic, I do not think it can be read as being “the” or even necessarily a very high part of Proto-Uralic ancestry. Rather, it is definitionally simply the genetics of the Khanty, which very likely wraps into it still also West Siberian ancestry components other than just Uralic. I would likewise also avoid readings such as 11 as “Proto-Samic”, 14 as “Proto-Samoyedic”, 18 as “Proto-Yeniseic”; especially given the case of 16, which could not possibly be any sort of either “Proto-Volgaic” or “Proto-Bulghar”. The method is in the end still one of sorting synchronic variation into components, even though it assumes a model of components historically mixing together in some people… and, it seems, not mixing in others.
What we could also find in ADMIXTURE charts is, however, an interesting sort of a low-resolution but high-dimensional presentation of PCA-like results. n-component mixtures form an n-dimensional vector space with positive values, normed to 100% ancestry (i.e. the positive-sign 1/2ⁿ-th of a n-sphere), which removes one dimension; while PCA graphs are two-dimensional slices of an in principle arbitrarily-many-dimensional and un-normed vector space. That is, 2 ADMIX components are equal to a linear space from 100% of one ancestry to 100% of another. 3 components would span an octant of a sphere, that we could flatten to a triangle plot. 4 components would span a topological tetrahedron, 5 components would span a topological 5-simplex, etc.
Tambets et al.’s data appendix provides also principal component projections up to N=5. This analysis has tighter geographic cutoffs and lacks e.g. Africans, South Indians and Chukotkans; so I will compare these with cases up to K=9 and ignore (the dimensions represented by) dark brown, tan (as distinct from orange) and olive (as distinct from green). This conveniently also brings us to the Uralic-like component 9 in the analysis.
- Principal component 1 is strongly west to east distributed, lowest in Basques and Sardinians, highest in Han and Evens. Could be fairly well derived from the ratio [2] of (navy + blue + green) to (magenta + orange + yellow), but also even better and more simple: the ratio of blue to yellow at the K=3 level of analysis, where Khantys are still mixed, not off in their own corner.
- PC 2 is roughly northwest to southwest, highest in Swedish Sami followed by Finnics & Baltics, lowest in Saudis and one Brahui. Only medium-high in Siberian Uralics and only medium-low in Han. Below K=9, a ratio of (navy + orange) to (blue + green + yellow) works fairly well for this, while magenta again fails to be useful. The K=4 split of just navy from blue would not match with this however, we do need orange.
- PC 3 is high in Nganasan, middling in Sami, low in Sardinians, Basques, Han. Ratio of orange to non-orange works decently here. This will be in some cases improved once magenta is split off from orange, in others not.
- PC 4 seems at first weird geographically: very high in Sardinians, Saudis, Nganasans, Evens, very low in Indians and kinda low in Han. But I believe the sign of PC levels is not actually significant of anything; some charts in the study also already show the same PC at inverted scales. If we do this, ratio of (yellow + green) to (blue + navy + orange + magenta) works tolerably (even if this would still put the Han higher up than Indians).
- PC 5 is high in Sami, Arabs and again one Brahui, fairly high also in Khantys; low in Caucasians, slightly low also in Yakuts. This by now has clearly no ADMIX fit for it: at K=6…9 Caucasians and Arabs look like a very similar green + blue mixture, and they remain not too far apart thruout the analysis.
It seems we have already reached the point where PCA analysis diverges from ADMIX. Presumably the reason is a greedy vs. non-greedy algorithm. Early principal components are not affected by calculating additional ones. Early ADMIX splits however frequently end up re-balanced, capturing only a smaller proportion of ancestry, once we try to fit a specific higher number of components. Both approaches may have their benefits.
It would also really seem, also from here, that some of the dimensions spanned by ADMIX components are not independent. This concerns mainly the localized components. Looking at the east-west axis, we can find some peoples, say, between tan and magenta (Selkups) or between magenta and light blue (Finnics, Kola Sami), but we really have nothing directly between tan and light blue. So this could be in effect treated as more of a single axis, where we have added a somewhat artificial “magenta pole” in the middle to split this in two. This is probably already reflected in how magenta appears basically in half from orange, in half from navy! There could be many other pairings too though, such as tan → orange → green, or orange → magenta → navy. Anything like this that would provide dimensionality reduction would be probably analytical progress, but how to best do this kind of component combination, or to avoid adding components that do this, wouldn’t be an easy mathematical task.
Lastly, there is a notable pattern in the PCA results that seems to be wholly unavailable in the ADMIX results: the same few specific peoples always end up on the borders of the charts! Generally these are at geographic remove. Saudis, Sami and Han as southern, northern and eastern outliers seem about as expected. But also: Nganasans have this property more strongly than Evenkis and Evens to their east; Basques and Sardinians have this property more strongly than Portuguese to their west or Cypriots to their (net) south. I suspect this demonstrates PCA is in fact better than ADMIX at providing hints of actually archaic rather than merely distinctively-mixed ancestry components, that might lie further beyond these peoples’ positions on the charts. We know from other studies by now that this is what happens with European hunter-gatherer archeogenetic data: they end up even lower on PC 1 than Basques and Sardinians. If there are some signals of completely distinct ancestry groups to be found among Uralic peoples, these are probably better represented by ADMIX components 11 (Sami) and 14 (Nganasan) than by 9 (Khanty) and 16 (Mari). But then the West European history also at least suggests such signals might be very old, shaped by many migrations and repopulations before the modern day, unlikely to have been widely around in pure form.
All my freestyle analysis here is possibly re-deriving understanding that is already known to genetics scholars and enthusiasts. Still, if my own experience is anything to go by, most of the rest of us have probably not seen anywhere near enough of this kind of basic how-to-read-the-results discussion: the primary genetics papers themselves are usually rather terse on this. If anyone is passing by and knows of other good, perhaps illustrated, guides about what do ADMIX or PCA results actually mean, do drop a note!
[1] How accidental are these colors, one wonders…
[2] Or perhaps, their difference? I’d have to check with actual data to see what would be the best arithmetic relationship between these scales.
Freelance reconstruction 
An exercise in (over?)reading ADMIXTUREs
Discussion in the previous post’s comments has come around to what we think we know and what we might be able to infer about the genetics of Uralic speakers. I have no direct data on this at my disposal (nor would I know off the cuff what to do with it if I did), but once again I find that already staring at published results seems to show a lot of details that would allow also closer discussion.
The most popular presentation of results on Uralic overall genetics might be still the 2018 paper “Genes reveal traces of common recent demographic history (…)” by the Tambets et al. group (surely it is by now the most influential, at least). This demonstrated, firstly, a broad west–east cline structure with offshoots to the south (Hungarian) and north (Samic), which may have been about as expected. But more interesting may have been their ADMIXTURE charts in Fig. 3A, that revealed at 9-component resolution a component (in magenta in their figure) that seems to correlate surprizingly well with Uralic-speaking peoples.
The match is still far from perfect. This component seems to be basically absent in Hungarians, while being well present in Volga and South Siberian Turkic peoples, outright a majority in Kets, and trickling marginally also into e.g. Russians, Swedes and Poles in the west (much less so in Latvians), Mongolians in the east. Its peak representation within the study is at >90% in Khanty & in a distinct subgroup of Mansi. Some Selkups also show a high proportion. Elsewhere it is mostly 30–40% though, usually in primary competition with another ancestry component (in navy) that is well spread thru Europe and reaches >90% in Latvians, almost as much in Lithuanians and Estonians. The lowest proportion of magenta in Uralic speakers occurs in Nganasans, whose main ancestry component at this resolution (in orange) is in common with all Altaic-speaking populations (> 80% in Yakuts, Evenkis, Evens), well present also in Yukaghirs and Kets and with a clear trace in Japanese and a subgroup of Han. All these complications have left me with the feeling that trying to ascribe this ancestry as being specifically spread by Proto-Uralic speakers may be too optimistic. Most pessimistically, it could represent instead an already pre-Uralic genetic cline, or assimilation to one — especially under current theories where Uralic spread (at least) to the west essentially within a metal trade network, thus perhaps as a trade language, at which point in time there may then not even have been such a thing as a single “Proto-Uralic-speaking people” anymore. Probably further back in time we could find something fitting this definition (and contrary to my last post, still most likely as the actual majority ancestors of at least one actual Uralic-speaking group), but if people A teach their language to people B who teach it to people C and D, who teach it to…, we might quickly end up chasing false leads about genetic markers.
Tambets et al. highlight also the 11-component resolution, where a new component appears (in light blue) that captures both the previously magenta + navy ancestry of the Swedish Sami, ending up at 98% among them!, overtakes magenta also among Kola Sami and Finnic speakers (but interestingly not North Russians) and appears in some 5%-ish proportions eastwards until the Mansi (and some 1%-ish trace presence all across Siberia). This remains without specific discussion in the paper, but has by now often come up in discussions with colleagues. My current understanding is that the ADMIXTURE algorithm in fact prefers to fit components to specific populations or individuals actually present in its data. And when a population has a distinct ancestry component not found in pure form (in this case we are likely dealing with the Paleo-Laplandic and/or Paleo-Lakelandic substrates in the Sami), it cannot reconstruct it. Instead a new component ends up representing this actual ancestry + a proportion of a reference group’s other compareable ancestry components, which then means even a small fragment of this “underlying” ancestry, together with other shared components, will now appear in larger size also elsewhere. So since the Sami still have some general-European ancestry, perhaps with drift or selection having emphasized some specific part of it, we find this reflected in a trickle of light blue also in e.g. Bulgarians or Hungarians. Since they also still have some eastern / Siberian ancestry, we find a similar trickle also in e.g. Udmurts (who have a good amount of general-European ancestry too) or Khanty. But then if this component is not directly representative of some old Proto-Sami or pre-Proto-Sami ancestry, and instead groups in also parts of other ancestry components that tend to occur together with one, why not also the same in other components? We also see that the equation 50% magenta + 50% navy = light blue, roughly valid for the Sami, does not hold in other populations. “Finnish light blue” is closer to one part magenta + two parts navy, and also leaves a little bit of magenta behind. “Veps light blue” looks more like the inverse: 60% magenta + 40% navy. “Komi light blue” and “Udmurt light blue” look to me to be taking in a dash of green, too. “Mari light blue”, looking closely, could involve a tiny reduction also in their yellow-coded ancestry (but also alternately, from the appearence of the light brown ancestry component in the chart).
Moreover I have not seen or heard it noted so far at all that their fuller data, provided as an appendix, includes also further turns of this sort. A very similar update seems to happen at the 14-component resolution: here a component in dark orange appears that captures the majority of the previously magenta and orange ancestry of the Nganasans. Other peoples who gain a large chunk of this end up being the Nenets and Selkups to the west, Evenkis, Evens and Oroqens to the east, and one specific Yakut individual and Koryak individual. Smaller amounts appear in e.g. Khanty, Tuvinians and in a sub-population of Dolgans, and a small sliver in all Japanese. Interestingly this component also, even at a low proportion, seems to completely replace light orange ancestry in several Uralic-speaking groups: besides the Nenets, who now end up as ca. 60% magenta, 40% dark orange, this also holds for the Khanty, Mansi and Komis. And also, not for Maris. Surely a sign that their orange ancestry is specifically recent from Turkic speakers, not ancient Siberian independently of general Uralic origin.
Should we assume this is, as I speculated might be the case for the Sami-affiliated component 11, in reflection of some lost substrate population that was closest of all to Nganasans? This I am much less sure of. The order in which components appear in ADMIXTURE results is clearly largest-first. Component 14 is found in only few peoples and is thus forced to be distinguished late from “general Altaic” ancestry. I must wonder is this might change if we pared the dataset down, making e.g. Nganasans a larger proportion of the data; and perhaps in a different order we’d find also find different split topology, such as 14 initially grouping with 9?
A second update to the component analysis of several Uralic-speaking groups is no. 16 in pink, likely representing no ancient ancestry at all but the formation by convergence of a specific secondary mixture: a “Volga region” component, appearing at 90% in Maris, very high also in Udmurts and Chuvashes. There is medium presence in Bashkirs, Tatars and Komi, smaller amounts appear in Russians, Mordvins and Mansi, and clear traces in all Turkic-speaking people and even in Poles. Interestingly, perhaps none in the Finnic peoples (but the appendix image is not high resolution enough for me to be sure of this).
Here we might move in the blog post from “reading” into “overreading”. Can we perhaps on closer looking work out what does the order and exact form of the splitting of ADMIXTURE components signify more specifically?
K=3. At this resolution we have three very broad components: a dark brown component that we might call sub-Saharan African or West African, at 100% in Mandenkas and Yorubas; a blue component at 100% in e.g. Basques and widely spread in Eurasia; and a yellow component at 100% in e.g. the Han and the Japanese. [1] Not highly useful for understanding anything about Uralic peoples: everyone on European side mostly blue with a small dash of yellow, Khanty and Samoyedics majority yellow with a clear piece of blue; though there are signs of other interesting patterns elsewhere, such as about 40% blue in Ethiopians, 5% brown in Georgians and Armenians.
K=4. This splits the blue component in two, into a broadly northern one in navy and a southern one keeping the medium-blue. No one ends up as purely either. Blue without navy appears, it seems to me, only in Ethiopians and Indians! The highest absolute amount of navy is in Estonians, Latvians, Russians at around 80%. Armenians and Georgians come out at the other end with 90% blue, 10% navy. The Swedish Sami, Nenets and Khanty (but no one further east) seem to end up with complete blue → navy conversion. So might e.g. Dolgans, Evenkis or Kets, but they have so little blue that I cannot be sure. Yakuts keep on a sliver of medium blue, Volga region Turkics follow median Uralic speakers in becoming majority-navy, other Turkic speakers maintain medium blue as the majority.
K=5. This splits off from yellow an orange component, that appears broadly across Siberia and/or Uralic speakers and in traces across Europe. The westernmost people who retain any consistent yellow are Bashkirs. Most Common Turkic speakers have a nearly even split. Chuvashes by contrast convert nearly all their yellow into orange, patterning here as western. Yellow remains in Indian control groups like Kannadigas. Han gain some 2% orange but remain the “type specimen” of yellow. The Japanese and a Han subgroup that also shows a small dash of blue and, later on, green (perhaps indicative of something like Manchu ancestry?) gain maybe 3–4%.
K=6. A light green component appears that is at its maximum in India, replacing both blue and orange here. Most Iranic peoples and neighbors (the Burusho and Brahui) go from about 90% blue / 10% orange (and a dash of other colors) to about 65% green / 35% blue, and seem to add back a trace amount of yellow. Green appears also in people from the Caucasus (at the expense of all of orange, blue and navy) and, less strongly, in Turkic and Uralic speakers. Further west, we see in Saudis, Yemenites and Palestinians a seemingly total conversion of both orange and navy, some decrease in blue too, for 15% green. Turks, Azeris, Kurds, “Iranians” (= Persians proper?) add an even more noticable 20% but still keep maybe 2-5% navy and orange. Orange traces in Europe mostly turn green, with the exception of Russians and Germans. In Moroccans and in Mzāb (a Berber group) navy now disappears, Sardinians go from 50% blue / 50% navy to 75% / 25%, and Basques from 33% / 66% to 50% / 50% — without any green appearing in any of these four. In Ethiopians a tiny sliver of green turns up. Altogether a confusing turn of components that seems to be mainly really about rebalancing something about the very vague blue / navy split.
By this point it seems navy is no more generally northern but rather northwestern in its distribution, still at its strongest in Balto-Slavic, Finnic and Samic speakers, with Germanics and Volga Region Uralics not far behind. I do not know what to make of its medium presence in Siberia and therearound, with fairly high navy : green rations in e.g. Kets, Selkups and Mongolis. Some individuals are probably though simply mixed half-Russian: there looks to be e.g. one Yukaghir, two Koryaks, at least three Chukchis who remain about 50% navy + medium-blue to the end of the analysis even after other components yet have been split off from these. One other “Chukchi” looks entirely Slavic; or possibly e.g. part Mordvin or German, but anyway not at all Siberian.
K=7. A tan “Beringian” component splits off from the general Siberian orange, at almost 100% in most Koryaks and Chukchi, a neat 50% also in the individuals with large blue and navy proportions, and remains definitive of these people until the end of the analysis. Only the Evens gain a noteworthy minority proportion, but the better-resolution data in the main paper shows that a small 1–2% sliver of this component still extends thruout Asia and into eastern and central Europe in e.g. Bulgarians or Poles or Finns. Not, however, into the Sami or Swedes.
K=8. The confusing green component splits further. The Indian half of it turns dark olive, at its highest proportion in “Malayans” (does this mean Malayalams? I would not guess Malays). A small portion of this appears in Iranians, Azeris, Turkmens and an even smaller trace in Syrians, Yemenese, Lezgins, but Near Easterners mostly keep light green. Pathans, Balochis etc. end up evently split or slightly more green than olive. In the Caucasus light green now becomes the majority over blue. Europe thru Siberia show only light green; which also seems to marginally increase at the expense of blue in e.g. French, Italians, Maris and also in Ethiopians. No effects on proportion of navy. An odd side effect might be also turning the sliver of eastern ancestry in Turks and Azeris now from orange to yellow, although the proportions of these across Siberia and Central Asia remain stable.
K=9. Enter the “Uralic” component, distributed as described above. This turns out to be mainly at the expense of orange and navy. Most Khantys remain 98% magenta from here to the end of the analysis. In peoples who were not included in the figure in the main paper, a small dash appears also in e.g. Iranians, Germans, Caucasians (Chechens and Lezgins in particular), Pathans, Burushos, and even Mzābs. In a few of these cases, detectable amounts of Proto-Uralic or even para-Uralic ancestry would be quite surprizing, I think.
K=10. A light brown component appears at the expense of primarily blue, secondarily dark brown, up to 80%+ in Mzābs. Replaces blue completely in Ethiopians and also replaces small traces of blue that up to now have been present in the third (and vague) deep African control group of “Bantus”. In Saudis this replaces dark brown completely but leaves a tiny sliver of blue. Outside the Near East this appears e.g. at 20% in Cypriots, 10% in Portuguese, Sardinians, Armenians (only a trace in Georgians), maybe <5% in Spaniards and across the Balkans. A few localized bits of dark brown in e.g. Sindhi or Balochi seem to remain. The new version of medium blue concurrently becomes more prevalent in West Europe at the expense of navy, rising e.g. from 25% to 50% in French, from 50% to 80% in Basque. Nothing changes in Siberia.
K=11. The light blue Swedish Sami-maximized component appears and remains at >98% in them until the end of the study. No apparent side effects beyond the Uralic/Nordic zone.
K=12. Light brown splits in two. The component that keeps the color grows to 100% in some Mzābs and remains there until the end of the study, is well-present also in Moroccans, and appears also in Portuguese and Spaniards (thus, at the western end of its range). Probably everywhere else previous light brown turns a pinkish brown. Balochi and Brahui seem to develop a larger proportion of brown, maybe at the expense of their previous blueish (either medium or navy) ancestry.
K=13. Olive splits in two and claims also some more of light green, whereever it was present. The component that keeps the olive color is now at 80%+ in Brahui (previously 60% green 40% olive) and becomes a more notable minority in e.g. Iranians, Tajiks and Lezgins. A light dirty green becomes the new South Indian majority and also ousts remaining yellow from here. The only clear inroads this makes beyond South Asia are into Pathans and Burusho, maybe a sliver in Tajiks and Uzbeks. Light green claims also yet more of blue from Georgians and Abkhazes, rising in them now to 90%.
K=14. The Nganasan-maximized component in dark orange, as noted above. As one side effect, the proportion of magenta in Kalmyks rises from marginal to maybe 10%, at the expense of light orange.
K=15. Another split from the generic-Siberian orange: a grey component that claims about 90% ancestry in Yakuts (aside from the one Nganasan-like, or even more closely, Even-like individual), 75% in Evenkis, 60% in Dolgans (who all now have about no orange); 50% in Yukaghirs and Evens (who retain 10% and 15% orange respectively); 5% in Buryats and Mongolians (who remain majority orange), traces in Turkic speakers further to the west. These colors, too, thus remain largely orthogonal to the linguistic division of Siberia. Light orange proportion in Han seems to rise slightly.
K=16. The Volga component in pink, as noted above, and yet another one orthogonal to linguistic classification. No side effects elsewhere that I can see on the proportions of magenta / green / navy that it replaces.
K=17. Some sort of a Levant component in dark grey enters, at 100% in some Palestinians but petering out to <5% in others; mostly thus replacing pink-brown and green. 5% also in Cypriots, only trace amounts in Jordanians and, oddly enough, Basques, Romanians and Gagauzes. I’m not sure if this has any presence anywhere else.
K=18. Here we gain a Sardinian component in purple, replacing almost all of their medium blue and also most of their pink-brown. Concurrently, Basque now becomes the unquestioned type specimen of medium-blue at maybe 98%, losing even its last traces of navy; its dark gray from the previous step seems to turn into purple, though! In the Middle East and southern Caucasus (Armenians, Georgians, Abkhazians) blue also mainly turns into purple. Italians, Romanians and Slavic speakers gain more purple than blue; more blue than purple at least in western Romance, Germanic, Iranic, Finnic, Mordvinic speakers and everyone in the North Caucasus, including the Turkophones (Balkars, Kumyks, Nogais).
K=19. We have one last change affecting Uralic speakers: a sky blue component appears at 80%+ in Ket and very variable among Selkups, anywhere from 90%+ to 10%. Smaller proportions appear mainly in South Siberian Turkic: Tuvans, Shors, Altais, Khakassians; as well as in Nenetses and in the non-Khanty-like part of Mansis (the Khantys still remain solid magenta). Possibly a dash in Uzbeks and Tajiks, though these appear by now as highly mixed populations with a little of almost every component (the former plurality orange, the latter plurality green + almost as much olive). Maybe same in Kyrgyzes and Kazakhs, who have a stronger orange plurality but still likewise a very diverse collection of components.
K=20. The last component of the study rearranges again the greenish colors: a new bright green hits a 90% among some of the Burushos, turns up in some proportion in most Iranics / Central Asians, has again a spot appearence among Lezgins and Chechens, and I think even at least one Syrian. Really nothing of interest for Uralic though.
My initial impression from this deep dive is not so much that I’ve definitely learned new things (though interestingly many known results of population genetics do appear here in a nutshell, such as Near Eastern connections of the Sardinians or the Ethiopians) as much as, first of all, that there would be likely benefits to re-doing this more detailed analysis with fewer “distractions”. Suppose we discard those peoples that end up generating their “own” components without much relevance for distinguishing Uralic speakers or their neighbors: that is the Mzābs and Moroccans (#12), South Indians (#8 and #13), Palestinians and Jordanians (#17), Burushos (#20), maybe one (but maybe not both) of the Sardinians & the Basques (#18). Do we still find the other components appearing in the same shape and same order? What six other components do we obtain instead if running the analysis to K=20?
We might also leave aside groups like the Japanese, Orcadians, Balkars, Abkhazians, Ossetians who don’t seem to add much of their own, to have Uralics and neighbors be a larger proportional part of the data (yes, Northwest Caucasians and Ossetians may have had contacts with migration-era Hungarians, but we can by now take it as known that modern Hungarians look genetically generically Slavic ~ Romanian). And since the algorithm seems to also look at the pre-given ethnicity bins, what would happen if we e.g. split in advance the Mansi or Kola Sami samples in their apparent subgroups? Or removed the apparent half-Russians from amongst many peoples of Siberia?
My second observation is that there’s a suspicious lack of any component that we could think of as even vaguely tracking most other larger language families: Indo-European (or its subgroups) or Turkic or Tungusic, without spilling in large percentage into their neighboring language families. Orange after the Chukotkan split in K=7 does maybe a tolerable job at tracking classical Ural–Altaic, but still keeps on also Yukaghir and Ket. And there is no hope at any of the resolutions here for telling apart e.g. an Abkhaz from a Balkar or a Persian from an Azeri. Perhaps the ability to still distinguish Finns and Swedes, or Khantys from everyone, is somewhat a happy accident really. Everywhere else Uralic-speaking people continue to look more similar to their neighbors than to their distant linguistic relatives.
Thirdly, towards the end of study we have also ended up with a fairly clear split of components into two sorts: those that have their “type specimens” and those that do not. Most of the former seem to be also very localized, though a few are not.
The impression I get from the big picture is that these “components” are thus unlikely to represent any sort of recent unique markers. They seem to come mainly from already Paleolithic genetic variation, sometimes probably eventually approximating unique “poles” (yellow, medium-blue, maybe tan) but more often probably representing specific proportions of mixture, especially if that has been later on well preserved in some peoples. So in particular, even if component 9 had something to do with the spread of Uralic, I do not think it can be read as being “the” or even necessarily a very high part of Proto-Uralic ancestry. Rather, it is definitionally simply the genetics of the Khanty, which very likely wraps into it still also West Siberian ancestry components other than just Uralic. I would likewise also avoid readings such as 11 as “Proto-Samic”, 14 as “Proto-Samoyedic”, 18 as “Proto-Yeniseic”; especially given the case of 16, which could not possibly be any sort of either “Proto-Volgaic” or “Proto-Bulghar”. The method is in the end still one of sorting synchronic variation into components, even though it assumes a model of components historically mixing together in some people… and, it seems, not mixing in others.
What we could also find in ADMIXTURE charts is, however, an interesting sort of a low-resolution but high-dimensional presentation of PCA-like results. n-component mixtures form an n-dimensional vector space with positive values, normed to 100% ancestry (i.e. the positive-sign 1/2ⁿ-th of a n-sphere), which removes one dimension; while PCA graphs are two-dimensional slices of an in principle arbitrarily-many-dimensional and un-normed vector space. That is, 2 ADMIX components are equal to a linear space from 100% of one ancestry to 100% of another. 3 components would span an octant of a sphere, that we could flatten to a triangle plot. 4 components would span a topological tetrahedron, 5 components would span a topological 5-simplex, etc.
Tambets et al.’s data appendix provides also principal component projections up to N=5. This analysis has tighter geographic cutoffs and lacks e.g. Africans, South Indians and Chukotkans; so I will compare these with cases up to K=9 and ignore (the dimensions represented by) dark brown, tan (as distinct from orange) and olive (as distinct from green). This conveniently also brings us to the Uralic-like component 9 in the analysis.
It seems we have already reached the point where PCA analysis diverges from ADMIX. Presumably the reason is a greedy vs. non-greedy algorithm. Early principal components are not affected by calculating additional ones. Early ADMIX splits however frequently end up re-balanced, capturing only a smaller proportion of ancestry, once we try to fit a specific higher number of components. Both approaches may have their benefits.
It would also really seem, also from here, that some of the dimensions spanned by ADMIX components are not independent. This concerns mainly the localized components. Looking at the east-west axis, we can find some peoples, say, between tan and magenta (Selkups) or between magenta and light blue (Finnics, Kola Sami), but we really have nothing directly between tan and light blue. So this could be in effect treated as more of a single axis, where we have added a somewhat artificial “magenta pole” in the middle to split this in two. This is probably already reflected in how magenta appears basically in half from orange, in half from navy! There could be many other pairings too though, such as tan → orange → green, or orange → magenta → navy. Anything like this that would provide dimensionality reduction would be probably analytical progress, but how to best do this kind of component combination, or to avoid adding components that do this, wouldn’t be an easy mathematical task.
Lastly, there is a notable pattern in the PCA results that seems to be wholly unavailable in the ADMIX results: the same few specific peoples always end up on the borders of the charts! Generally these are at geographic remove. Saudis, Sami and Han as southern, northern and eastern outliers seem about as expected. But also: Nganasans have this property more strongly than Evenkis and Evens to their east; Basques and Sardinians have this property more strongly than Portuguese to their west or Cypriots to their (net) south. I suspect this demonstrates PCA is in fact better than ADMIX at providing hints of actually archaic rather than merely distinctively-mixed ancestry components, that might lie further beyond these peoples’ positions on the charts. We know from other studies by now that this is what happens with European hunter-gatherer archeogenetic data: they end up even lower on PC 1 than Basques and Sardinians. If there are some signals of completely distinct ancestry groups to be found among Uralic peoples, these are probably better represented by ADMIX components 11 (Sami) and 14 (Nganasan) than by 9 (Khanty) and 16 (Mari). But then the West European history also at least suggests such signals might be very old, shaped by many migrations and repopulations before the modern day, unlikely to have been widely around in pure form.
All my freestyle analysis here is possibly re-deriving understanding that is already known to genetics scholars and enthusiasts. Still, if my own experience is anything to go by, most of the rest of us have probably not seen anywhere near enough of this kind of basic how-to-read-the-results discussion: the primary genetics papers themselves are usually rather terse on this. If anyone is passing by and knows of other good, perhaps illustrated, guides about what do ADMIX or PCA results actually mean, do drop a note!
[1] How accidental are these colors, one wonders…
[2] Or perhaps, their difference? I’d have to check with actual data to see what would be the best arithmetic relationship between these scales.
Posted in Commentary, Methodology